P. 1
US Federal Reserve: Paper-Wright

US Federal Reserve: Paper-Wright

|Views: 15|Likes:
Published by The Fed

More info:

Published by: The Fed on Jan 23, 2008
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

03/04/2015

pdf

text

original

Comparing Greenbook and Reduced Form Forecasts

using a Large Realtime Dataset

Jon Faust
y
Jonathan H. Wright
z

We thank Douglas Battenberg, Jean Boivin, Bryan Campbell, Frank Diebold, Mike McCracken,
Athanasios Orphanides, Lucrezia Reichlin, Dave Reifschneider, Chris Sims, Jim Stock and Mark Watson
for helpful comments. All remaining errors are our own. The views expressed in this paper are solely
the responsibility of the authors and should not be interpreted as re‡ecting the views of the Board of
Governors of the Federal Reserve System or of any other employee of the Federal Reserve System.
y
Department of Economics, Johns Hopkins University, Baltimore MD 21218 and Division of In-
ternational Finance, Federal Reserve Board, Washington DC 20551, phone: (202) 452-2328, e-mail:
faustj@jhu.edu
z
Division of Monetary A¤airs, Federal Reserve Board, Washington DC 20551, phone: (202) 452-3605,
e-mail: jonathan.h.wright@frb.gov.
Abstract: Many recent papers have found that atheoretical forecasting methods using
many predictors give better predictions for key macroeconomic variables than various
small-model methods. The practical relevance of these results is open to question, how-
ever, because these papers generally use ex-post revised data not available to forecasters
and because no comparison is made to best actual practice. We provide some evidence
on both of these points using a new large dataset of vintage data synchronized with
the Fed’s Greenbook forecast. This dataset consists of a large number of variables, as
observed at the time of each Greenbook forecast since 1979. Thus, we can compare real-
time large dataset predictions to both simple univariate methods and to the Greenbook
forecast. For in‡ation we …nd that univariate methods are dominated by the best atheo-
retical large dataset methods and that these, in turn, are dominated by Greenbook. For
GDP growth, in contrast, we …nd that once one takes account of Greenbook’s advantage
in evaluating the current state of the economy, neither large dataset methods nor the
Greenbook process o¤ers much advantage over a univariate autoregressive forecast.
1. Introduction
In recent years, researchers have investigated many di¤erent ways of forecasting an eco-
nomic time series using a large number of predictors—say, 40 or more (e.g., Stock and
Watson, 1999, 2002, 2003, 2005; Bernanke, Boivin and Eliasz, 2005; Giannone, Reichlin
and Sala, 2004; Forni, Hallin, Lippi and Reichlin, 2005). When the number of predictors
is large relative to the available sample size, one must constrain the estimation of the
forecasting model in some way in order to avoid the perils of over…tting, and the many
methods di¤er mainly in how they do so. The methods have one aspect in common,
however: they are atheoretical in the sense that restrictions from economic theory are
eschewed.
1
A consistent result in this work is that these large-dataset methods outperform
various naive and semi-sophisticated benchmarks including random walk forecasts, simple
univariate time series models, and, in some cases, simple models motivated by economic
theory.
While these pioneering results are tantalizing, their importance for practical fore-
casting is di¢cult to assess for two related reasons. First, arguably none of the bench-
marks used in this research is used for practical forecasting. While the simple bench-
marks are probably the right starting point for assessing new methods, ultimately we care
whether the new methods outperform standard practice or best practice in forecasting.
Second, comparison to real-world forecasting methods is complicated by the issue of data
revisions. Many macro time series are heavily revised through time. Real-world forecasts
use the noisy early vintages available in real-time; large-dataset forecasting research has
been conducted almost exclusively with a single vintage of revised data.
Bernanke and Boivin (2003) provide a notable exception. They use a real-time
dataset to assess various large dataset forecasting methods and they include a comparison
to the Fed’s Greenbook forecast. Two results are especially interesting and provocative.
1
More speci…cally, economics only guides the selection of variables to include in the exercise.
1
First, factor models generally predict less well in their 78-variable real-time dataset than
using the 215 series of fully-revised data used by Stock and Watson (2002). This appears
to owe mainly to the larger size of the Stock and Watson dataset, because the factor
models also predict less well in the revised version of the 78-variable dataset. Second,
the Greenbook seems to outperform all the other methods considered—regardless of the
dataset employed or whether revised data are used.
In this paper, we take up some of the questions raised by Bernanke and Boivin
(2003), using a unique set of vintage data associated with the Federal Reserve’s Green-
book forecast, which is prepared for each FOMC meeting.
2
The dataset has a snapshot
of a large number of macroeconomic time series as they existed at the time of 145 Green-
books between March 1980 and December 2000. These data allow us to create pseudo-
real-time large dataset forecasts using information sets that are precisely synchronized
with the Greenbook.
First, we compare various large dataset forecasts of output growth and in‡ation to
simple benchmarks in real time, to see if the conclusions are sensitive to data-vintage
issues. Following Bernanke and Boivin (2003), we study factor model approaches. We
also study two model averaging methods in which the overall forecast is a weighted
average of forecasts from a large number of simple bivariate models. We …nd that the
model averaging methods are both more robust and generally more accurate than the
factor model methods—the factor models sometimes perform very badly.
Second, we compare the large dataset forecasts to the Greenbook forecasts, exploit-
ing the fact that the information sets are perfectly synchronized. This comparison is
particularly interesting because much earlier work suggests that the Greenbook forecast
2
Other authors who have used these electronic archives in this way include Tetlow and Ironside (2004),
who evaluated alternative policy rules with real-time vintages of the data and of the FRB/US model,
and Edge, Kiley and Laforte (2006), who used data on 11 series from these databases since 1996 to
compare Greenbook forecasts with real-time forecasts based on Bayesian VARs and dynamic stochastic
general equilibrium (DSGE) models.
2
has generally been at or near the frontier of best performance in forecasting. We are not
aware of documentation of a real-time forecast that consistently outperformed Greenbook
over this period.
3
We do not mean to imply that Greenbook is known to be optimal in
some sense, only that it is near the frontier of best actual practice, and thereby an inter-
esting benchmark. Further, the Greenbook is a subjective forecast based on an immense
range of information processed through a subjective …lter that is guided by economics.
4
Thus, we can view the large-dataset-forecast versus Greenbook comparison as a test of
the atheoretical use of a large dataset versus sophisticated use of an immense dataset.
Using this unique dataset, we can considerably sharpen the comparison of Green-
book to other methods over what has been done heretofore. In particular, we can evaluate
whether the well-documented Greenbook advantages stemmainly fromsuperior estimates
of the state of the economy at the time of the forecast. When the Greenbook forecast for
GDP is made in quarter t, very little data for quarter t may be available, and some data
for quarter t1 will not be available. Sims (2002) has suggested that the good properties
of Greenbook might ‡ow from the great e¤ort the Fed makes to evaluate the current or
recent past state of the economy at the time of the forecast. For example, by mirroring
key elements of the data construction machinery of the Bureau of Economic Analysis,
the Fed sta¤ can form a relatively precise estimate of what BEA will announce for the
previous quarter’s GDP even before it is announced. Further, the sta¤ can adjust the
estimate of the current state for certain large transitory events such as dock strikes, hur-
ricanes, etc. Given the persistence in economic measures, a better estimate of persistent
and transitory components of the current state may translate into forecasting advantages
over moderate horizons.
To put is most starkly, one might conjecture that the Fed’s forecasting advantage
3
Romer and Romer (2000) found that Greenbook outperformed private sector surveys and Sims
(2002) found generally favorable results, especially for Greenbook in‡ation forecasts. Bernanke and
Boivin (2003) reach the same conclusion.
4
For a discussion of the Greenbook forecasting process, see Reifschneider, Stockton and Wilcox (1997).
3
stems purely from measuring the current state, with little or no advantage over atheoret-
ical methods in projecting what the current state implies for the future. We assess this
conjecture in a simple way. Consider the vintage data available in quarter t, so that the
released values of many data series end in quarter t 1 or t 2. We have a Fed forecast
of each variable, so we can append the forecast to the actual data, and bring the data
all up to quarter t 1, or t, or t + ,. We can then use this updated dataset to form the
various time series forecasts. We call the point we update the data to “the jumping-o¤
point” for the time series methods and consider jumping-o¤ points from quarter t 1
through t + 3. If the Fed’s advantage mainly ‡ows from assessing time t 1 or t, the
atheoretical methods jumping o¤ from time t might compare favorably to Greenbook.
Given the generally favorable existing results on Greenbook, our initial suspicion
was that for both in‡ation and output growth, Greenbook would look good at very
short horizons, but that the time series methods jumping o¤ from some point a quarter
or two in the future would compete favorably. This suspicion turns out to be wrong
for both in‡ation and output, but, somewhat surprisingly, in di¤erent directions. For
in‡ation, Greenbook generally outperforms other methods (often substantially) for all
jumping o¤ points out to t + 3: even when the time series models know what the Fed
is thinking about in‡ation three quarters into the future, the atheoretical methods are
dominated by Greenbook in forecasting further into the future. In sharp contrast, for
output growth, once we give the time series methods quarter t, the current state, the
Greenbook advantage disappears. Perhaps more surprisingly, none of the time series
methods clearly outperforms an AR(4) predicting in output growth, using jumping-o¤
points of quarter t or later.
Overall, then, for in‡ation and for all jumping-o¤ points, Greenbook dominates the
best large model methods, which dominate the small model methods. For output, for
all jumping-o¤ points except t 1, the best atheoretical methods and Greenbook per-
form comparably and no method clearly dominates the naive benchmark of a univariate
4
autoregression. Knowledge of interest rates, in‡ation, unemployment, etc., do not allow
one to improve on the forecast that growth will revert at the historical rate back to the
current estimate of the mean.
The plan for the remainder of this paper is as follows. Sections 2 and 3 describe
our real-time dataset and methods, respectively. Section 4 contains the main results;
section 5 addresses a number of additional topics. Section 6 concludes.
2. The Real-Time Dataset
Before each FOMC meeting, the Federal Reserve sta¤ prepares a brie…ng document
called the Greenbook, which contains a sta¤ forecast of the macroeconomy. While the
Greenbook forecast is subjective, a large-scale econometric model has long been one tool
used in the Greenbook process. Since 1996, the FRB/US model has been used. Before
that, a model called variously the FMP (Federal Reserve-MIT-Penn) or MPS (MIT-Penn-
SSRC) was employed. Since 1979, the model database has been archived electronically
on the date that the Greenbook is published, a few days before the FOMC meeting.
These archived databases contain data on many variables of which some (but not all) are
forecasted in the Greenbook. The databases contain the historical sample of each variable
in the model as observed at that time, and forecasts. While the forecasts are generated
from the econometric model, an add-factor is then included to force the forecasts to
match those in the Greenbook, for those variables that are forecasted in the Greenbook.
Thus these archives comprise a large vintage dataset that is perfectly synchronized with
the Greenbook forecast itself.
We extracted 145 vintages of data from these electronic archives covering Green-
books from March 1980 to December 2000. The data used in this paper stop in 2000
because it is Federal Reserve policy that the Greenbook forecast cannot be released until
5 years after the forecast date. A few vintages are lost or never existed, including those
during early 1996 during the transition from the MPS model to the FRB/US model. The
5
maximum forecast horizon varies from one Greenbook to the next, but we considered only
Greenbooks for which the forecast horizon goes out to at least 5 quarters. Table 1 gives
the publication dates for our 145 vintages of data.
Ideally, we might like to have the same forecast variables in each vintage. Unfortu-
nately, the variables available in any given vintage vary a good bit. There is a large break
in the list of available variables at the time of the major model revision in 1996. Finally,
the amount of historical data for some series varies a bit across vintages. Further, the
nature of the model databases presents an additional variable selection problem. Many
variables in the databases are constructed from other variables based on identities, other
transforms of included variables, and slightly di¤erent versions of included variables (e.g.,
foreign sector variables on both the NIPA and BOP basis).
We extracted 112 macro time series, listed in Table 2, from these databases. We
omitted variables that are constructed from other variables and di¤erent de…nitions of
the same variable. From each vintage, we keep any of these 112 that have historical
data back to 1960Q2. We augment each vintage with the dividend-exclusive returns on
the S&P500 stock index.
5
In total, the number of series in each vintage ranges from 44 to 78, with an average
of 64. This is a large number of predictors, although it falls short of the very large datasets
of more than 100 variables that are used in some papers. Evidence on the marginal gain
in predictive power from increasing the number of variables from (say) about 50 to over
100 is mixed. Boivin and Ng (2005) …nd little marginal gain in predictive power, but
Bernanke and Boivin (2003) …nd a notable gain in predictive power from using a very
large dataset.
In the end, we have 145 vintages containing a varying list of variables, each of which
has history back to 1960Q2 and for which we have 5 quarter forecasts. The vast majority
5
To each vintage we added only observations that would have been available at the Greenbook
publication date. These data are of course not subject to revision
6
of our predictors are available out to the Greenbook forecast horizon. As is usual, we
transformed many of the series so that the transformed series is arguably stationary.
Some series are kept in levels (no transformation), others are in di¤erences and others
are in log di¤erences. Table 2 notes the transformation that was applied to each series.
The three series that we forecast in this paper are the quarterly in‡ation rate as
measured by the CPI, the quarterly in‡ation rate as measured by the GNP/GDP de‡ator
and the quarterly real growth rate (GNP/GDP). As is standard, we use GNP during
and before 1991, and GDP subsequently. The in‡ation rate and the growth rate are
computed as 400 log(r
t
r
t1
) where r is the price or output series.
Our goal is to assess di¤erent methods for forecasting these in‡ation and growth
variables in real-time. But an important issue that arises is that these variables are in
turn constantly being revised and so it is not clear what to treat as the realized data.
For the national income and product accounts, the source of our forecast variables, the
…rst data release (known as the advance release) comes out about one month after the
end of the quarter to which the data refer. The data are then revised in the preliminary
and …nal releases, which incorporate more source data, and are released about two and
three months after the end of the quarter to which the data refer, respectively. The data
then continue to get revised inde…nitely, through annual and then benchmark revisions,
with the latter incorporating conceptual and de…nitional changes. Since our objective
is the comparison of Greenbook and statistical forecasts, and since Federal Reserve sta¤
attempt to forecast variables as they are de…ned at the time that the forecast is made,
we measure actual realized in‡ation and growth by the data as recorded in the real-time
dataset of the Federal Reserve Bank of Philadelphia two quarters after the quarter to
which the data refer. This will typically correspond to the data as recorded in the second
revision of the national income and product accounts.
7
3. Methods
3.1 The forecasting models
We construct forecasts using 10 time series methods from the literature. Call the variable
to be forecast n
t
(a measure of in‡ation or output growth) and a collection of potential
predictors fr
it
g
n
i=1
. At time 1 we observe data up to some point in time 1 /. We will
consider forecasts for n
T+h
, / = 0. . . . . 5.
We consider the following time series forecasts:
1. A random walk model (RW). This takes n
T
as the forecast for n
T+h
. This is close to,
but not quite the same as, the forecast for in‡ation considered by Atkeson and Ohanian
(2001), who take
1
4

3
j=0
n
Tj
as the forecast for n
T+h
.
2. Iterated autoregression (IAR). We estimate n
t
= j
0
+
p
j=1
j
j
n
tj
+
t
(we use j = 4).
The /-period forecast is constructed by iterating the one-step forecast forward.
3. Direct forecast from autoregression (DAR). For each /, we estimate n
t+h
= j
0
+

p
j=1
j
j
n
tj
+
t
(we use j = 4). Each /-step forecast is a one-step forecast from the
model for the appropriate /. The iterated AR forecast will asymptotically outperform
the direct model if the AR(4) model is correctly speci…ed, but the direct forecast may be
more robust to misspeci…cation, as discussed by Marcellino, Stock and Watson (2006).
4. Equal weighted averaging (EWA). We …rst estimate and forecast using : simple
models, each of the form n
t+h
= j
0
+
p
j=1
j
j
n
tj
+
i
r
it
+
it
for i = 1. . . . : (we use
j = 4). Letting ^ n
i
T+h
be the forecast of n
T+h
from the ith model, the EWA forecast
is :
1

n
i=1
^ n
i
T+h
. This method was …rst proposed by Bates and Granger (1969) and its
empirical success is part of the folklore of forecasting. The method is considered by Stock
and Watson (2003), among others.
5. Bayesian model averaging (BMA). In this method, described in more detail by Wright
(2003), we take the : models as in EWA and form a weighted average forecast motivated
by the following Bayesian reasoning. Assign a prior in which each model is equally likely
to be the true model. For the prior for the model parameters, we follow Fernandez, Ley
8
and Steel (2001). Write each model as n
t
= \
0
i
n
it
+
it
. Assume that
it
`(0. o
2
) and
assume that the prior for \
i
conditional on o is `(

\. c(o
2

n
i=1
n
it
n
0
it
)
1
) ;

\ is described
below. The (improper) marginal prior for o is proportional to 1o. In this case, the
posterior mean of \
i
is
~
\
i
=
^

i

1
+

1
where
^
\
i
denotes the OLS estimate of \
i
while
the probability that the i
th
model is the true model is
1(`
i
) = (
1
1+
)
k=2
[() \
i

\)
0
() \
i

\)

1+
() \
i

\)
0
\
i
(\
0
i
\
i
)
1
\
0
i
() \
i

\)]
T=2
where ) = (n
1
. n
2
. ...n
T
)
0
and \
i
= (n
i1
. n
i2
. ...n
iT
)
0
. Accordingly, the proposed forecast
is
n
i=1
1(`
i
)
~
\
0
i
n
it
. In our implementation, we will set j = 4, and let

\ = ( j
0
. j
1
. ... j
p
.

i
)
0
where

i
= 0 and j
0
. j
1
. ... j
p
are obtained from …tting an autoregression to the pre-sample
consisting of the data for 1947Q1 to 1960Q1 as recorded in the 1978Q4 vintage of the
Federal Reserve Bank of Philadelphia real-time dataset.
The theoretical justi…cation of this method relies on strictly exogenous regressors
and iid errors—assumptions that are patently false in our application. Earlier work
(Koop and Potter (2003) and Wright (2003)) shows that the method works well in cases
like the one at hand, however, and we simply view BMA as a pragmatic shrinkage device.
6. Factor augmented autoregression (FAAR): For each /, we estimate n
t+h
= j
0
+

p
j=1
j
j
n
tj
+
m
i=1

i
.
it
+
t
where f.
it
g
m
i=1
are the …rst : principal components of fr
it
g
n
i=1
.
The predictors are …rst standardized to have mean zero and unit variance. We use j = 4,
: = 3. The forecasts are then constructed as in the direct AR forecast.
7. Factor augmented vector autoregression (FAV): :
t
= j
0
+
p
j=1
j
j
:
tj
+
t
, where
:
t
= (n
t
. .
1t
. .
2t
. ....
mt
)
0
. We set j = 1 and : = 3. The model can be estimated and
iterated forward to provide a forecast of n
T+h
. This method was proposed by Bernanke,
Boivin and Eliasz (2005).
8. An integrated factor augmented VAR (IFAV). :
t
= j
0
+
p
j=1
j
j
:
tj
+
t
, where
:
t
= (n
t
. .
1t
. .
2t
. ....
mt
)
0
. We set j = 1 and : = 3. The model can be estimated and
9
iterated forward to provide a forecast of n
T+h
. This is the same as 4, except that a unit
root is being imposed in the variable to be forecast.
9. A dynamic factor model (DF). This method, described in detail by Forni, Hallin,
Lippi and Reichlin (2005), takes the vector of the stochastic standardized components of
the data, .
t
, (both the variable being forecast and the predictors), and assumes that is
can be represented as .
t
= .
t
+.
t
where .
t
= 1(1)1
t
is a vector of common components,
.
t
is a vector of idiosyncratic components, 1
t
is a ¡x1 vector of dynamic factors that follow
a stationary vector autoregression of order :, and the processes .
t
and .
t
are mutually
orthogonal at all leads and lags. The identifying assumption is that the eigenvalues of
the spectral density matrix of .
t
are uniformly bounded in the limit as the number of
predictors gets large, whereas the eigenvalues of the spectral density matrix of .
t
are
unbounded as the number of predictors gets large. The model can be written in the
form .
t
= (1
t
+ .
t
where 1
t
= (1
0
t
. 1
0
t1
...1
0
ts
)
0
is a vector of : = ¡(: + 1) static factors.
Given an estimator of the spectral density matrix of .
t
, Forni, Hallin, Lippi and Reichlin
show how to use dynamic principal components analysis to estimate the spectral density
matrices of .
t
and .
t
and hence to construct /-step ahead forecasts of .
t
. Following
Forni, Hallin, Lippi and Reichlin and also D’Agostino and Giannone (2006), we use this
as the /-step ahead forecast for .
t
, treating .
t
as white noise for prediction purposes.
The method di¤ers from standard principal components analysis in that the dynamic
restrictions on 1
t
are imposed
6
and the di¤erent series are weighted by their respective
signal-to-noise ratios in estimating the factors. We set ¡ = 3, : = 15 and estimate the
spectral density matrix of .
t
using a Bartlett window with truncation lag equal to the
square root of the sample size.
7
10. An unobserved component stochastic volatility model (UCSV). Stock and Watson
(2007) …nd that an unobserved components model with stochastic volatility provides
6
Concretely, the restriction is that the spectral density matrix of F
t
must have rank q.
7
We are grateful to Mario Forni for providing us with the code for implementing this procedure.
10
good forecasts for in‡ation. The model is a univariate speci…cation that n
t
= t
t
+:
T
t
and
t
t
= t
t1
+:
P
t
where :
T
t
is iid`(0. o
2
T;t
), :
P
t
is iid`(0. o
2
P;t
), log(o
2
T;t
) = log(o
2
T;t1
) +·
1;t
,
log(o
2
P;t
) = log(o
2
P;t1
) + ·
2;t
and (·
1;t
. ·
2;t
)
0
is iid`(0. 1
2
). The interpretation of the
model is that n
t
is the sum of an stochastic trend and noise, with both the volatility of the
noise (temporary shocks) and the shocks to the stochastic trend (permanent shocks) being
time-varying. The model implies that the series is I(1). Stock and Watson found that
this model forecast in‡ation well because it recognized that the permanent component
of in‡ation had high variance in the early 1980s, but became more stable subsequently.
The model can be estimated by Markov Chain Monte Carlo. The forecast of n
T+h
from
this model is simply the …ltered estimate of t
T
.
Finally, we note that throughout this paper we consider forecasting of one-quarter
growth or in‡ation, / quarters hence. Many authors instead consider the prediction of
cumulative growth of in‡ation from quarter t 1 to quarter t +/, or four-quarter growth
or in‡ation ending / quarters hence. These are of course all di¤erent ways of expressing
the same information. However, since one of our purposes in this paper is to assess
the relative information content and Greenbook forecasts at di¤erent horizons, it seemed
best to us to report results in terms of one-quarter growth at di¤erent horizons, as these
other measures confound short- and longer-term predictive ability.
3.2 Bootstrap p-values for non-nested comparisons
One of our goals is to compare the forecast accuracy of the Greenbook with the time
series methods. The appropriate construction of p-values for a comparison of root mean
square prediction errors depends on whether the forecasts being compared are nested, or
non-nested. We think it is appropriate to assume that the time series forecasts are not
nested in the Greenbook. We accordingly approximate the sampling distribution of the
RMSPEs using the moving-blocks bootstrap of Künsch (1989) and Liu and Singh (1992).
8
8
We use a block length of 10, corresponding to a span of a bit more than one year of data.
11
This is e¤ectively using the moving-blocks bootstrap to simulate the distribution of the
Diebold and Mariano (1995) statistic.
9
We would also like to test the relative merits of the di¤erent atheoretical time series
methods. But many pairs of forecasting models that we consider are nested. The test
statistic of Diebold and Mariano (1995) has a nonstandard distribution under the null
hypothesis of equal forecast accuracy if the models are nested, as the models are the same
under the null (Clark and McCracken (2001)), and the bootstrap is not valid even to …rst
order. Thus our bootstrap p-values do not allow us to make nested model comparisons.
4. Results
Our main metric for forecast accuracy is the root mean square prediction error (RMSPE)
for the various forecasting methods (Table 3). We provide RMSPEs at each horizon 0
through 5—that is, the current quarter through 5 quarters hence. For each of the time
series forecasting methods, we provide RMSPEs of forecasts that condition on data up to
and including quarter t 1, where quarter t denotes the quarter in which the Greenbook
was published. We also report the results of the time series forecasts that take later
jumping-o¤ points, i.e. that condition on data up to and including quarter t, t + 1,
t + 2 or t + 3, using Greenbook forecasts for current and future quarters. Presuming
that Greenbook is best at very short horizons, this method allows us to ask if there is
some horizon at which one can usefully switch from Greenbook to one of the atheoretical
methods.
When the RMSPE for a time series forecasting model is less than the corresponding
value for Greenbook, the entry is in bold. Cases in which the RMSPE is signi…cantly
di¤erent from the corresponding value for Greenbook are marked at 1, 5 and 10 percent
signi…cance levels, using the two-tailed Diebold-Mariano test with bootstrap j-values as
9
The asymptotic p-values for the Diebold-Mariano test using Newey-West standard errors with a lag
length of 10 are not shown, but are similar to those implied by the bootstrap.
12
described above.
4.1 Greenbook versus atheoretical models large and small
The results for in‡ation and output growth are strikingly di¤erent. For in‡ation (either
the de‡ator or CPI in‡ation), a quick summary goes like this. Greenbook dominates
all the time series methods at all forecasting horizons and for almost all jumping-o¤
points.
10
The point estimates for Greenbook are typically 10 to 40 percent smaller than
those of the other models and we can reject the hypothesis of equal forecast accuracy of
Greenbook and the other methods.
The main exception to this general result is that as we move the jumping-o¤ point
out in time, we cannot reject the hypothesis that Bayesian model averaging is as accurate
as Greenbook. However, the point estimates for Greenbook remain smaller than those
for BMA, except for the farthest out jumping-o¤ points and horizons.
For output growth, Greenbook is a good deal better than the atheoretical methods
at horizon zero. This is consistent with the view that the Fed usefully exploits a great
deal more information about the current state of the economy than is used in the time
series models. After horizon zero, however, the advantage largely evaporates. One can
see this in two ways. First, if we keep the jumping-o¤ point at t 1, the point estimates
of Greenbook and the other models are comparable as we consider horizons beyond zero.
Thus, the Greenbook advantage at time zero does not translate into forecasting gains
at other horizons. Second, if we move the jumping-o¤ point out even one quarter, the
Greenbook advantage at all remaining horizons disappears, and is perhaps even reversed.
That is, for several time series methods the point estimate of the RMSPE is smaller and
we cannot reject the hypothesis that the method is at least as good as Greenbook.
As another metric for comparing the Greenbook and atheoretical forecasts, Table 4
10
There appear to be quite strong seasonal patterns in many vintages of the GDP/GNP de‡ator
in‡ation data, despite the fact that these data are seasonally adjusted. We experimented with adding
deterministic seasonal dummies to each of the forecasting methods, but found that this nearly always
increased mean-square prediction errors.
13
shows the proportion of forecasts for which each atheoretical model is more accurate ex-
post than the Greenbook. Thus, numbers greater than 50 percent favor the atheoretical
forecasts, while numbers smaller than 50 percent favor the Greenbook. For in‡ation,
Greenbook does generally does better than the time series forecasts a good bit more
than half the time. An exception is that BMA does better than Greenbook a bit more
than half the time in forecasting CPI in‡ation at longer horizons. For growth, Greenbook
does better than the time series methods well more than half the time at horizon zero,
but at all other horizons Greenbook and time series models seem about equally likely to
be more accurate.
These results suggest a signi…cant di¤erence in the forecastability of output and
in‡ation. This result becomes even more distinct in the next section.
4.2 Comparison of atheoretical methods
We are also interested in evaluating the relative merits of large dataset methods versus
univariate methods. For in‡ation (either de‡ator or CPI), among the univariate forecasts,
the iterated autoregression and UCSV model generally seem to do best. However,
Bayesian model averaging does a good bit better than any of the univariate in‡ation
forecasts, and generally gives the smallest RMSPE among all the atheoretical in‡ation
forecasts considered. The factor-augmented vector autoregression and dynamic factor
forecasts are notably less accurate predictors of in‡ation than the iterated autoregression
and UCSV forecasts. However, the factor augmented-vector autoregression that imposes
a unit root in in‡ation (IFAV) does better than the univariate forecasts, though still not
quite as well as BMA.
For growth, among the univariate forecasts, the iterated autoregression generally
gives the smallest RMSPE, and the univariate forecasts that impose a unit root (ran-
dom walk and UCSV) predict poorly. The forecasts based on equal-weighted averaging,
Bayesian model averaging and the factor-augmented vector autoregression all typically
14
give smaller RMSPE than the iterated AR forecast, but the gains are modest. Even this
advantage, however, largely disappears when we consider only the period since what has
become known as the Great Moderation, as discussed below.
4.3 Comments
One striking observation from Table 3 is how small the Greenbook prediction errors for
in‡ation are relative to any of the other forecasting procedures. This is consistent with
Ang, Bekaert and Wei (2006) who …nd that private sector surveys outperform both uni-
variate and multivariate time series forecasts of in‡ation. Ang, Bekaert and Wei make a
conjecture similar to that of Sims (2002) regarding the Greenbook: the subjective meth-
ods may be able to aggregate very diverse information and to adapt rapidly to changes
or special circumstances. These results sharpen those of Ang, Bekaert and Wei, however,
in showing that the advantage of the subjective methods over atheoretical methods in
forecasting in‡ation remains even when the atheoretical methods take advantage of the
subjective assessment up through three quarters in the future.
11
Whether this advan-
tage of Greenbook stems from access to a greater range of information or from a more
sophisticated use of the available information, or both, remains an open question.
Our results show a clear pattern favoring the model averaging methods over the
factor model approaches. Both model averaging approaches are consistently among the
best in the results so far, with Bayesian model averaging perhaps having a slight edge
over equal-weighted averaging. This result remains throughout the robustness checks
performed below. Some factor model methods perform very badly in some cases and
none of these methods consistently performs as well as the Bayesian model averaging
approach. The fact that the factor model methods sometimes fare poorly was also noted
by Bernanke and Boivin (2003). As they note, factor models fare better using the larger
11
It would be natural to compare the Greenbook forecasts in this dataset with those from private
sector forecasts. We leave this for future work, but note that this comparison is complicated by timing
di¤erences between the Greenbook and private sector forecasts.
15
dataset of Stock and Watson (2002). We take up this issue further below.
Notice further that while the Greenbook RMSPE pro…le increases with horizon for
de‡ator in‡ation, the pro…le is essentially ‡at from horizon 1 onward for CPI growth
and output growth. Thus, in‡ation measured by the GDP de‡ator becomes increasingly
di¢cult to forecast as the horizon increases, and more data and subjective methods both
seem to help. In contrast, GDP growth remains equally forecastable over horizons 1
through 5 and there are no clear gains from using more data or the subjective methods
from Greenbook.
5. Additional topics
In this section, we take up several additional topics that shed light on the robustness of
the results.
5.1 Sub-samples and the Great Moderation
The period from 1979–1983 was especially volatile in the U.S. economy, containing the
sharpest and deepest recessions in the post-War era. The period since 1982 has seen the
Great Moderation in which the economy seems much less volatile even than the period
before 1979. Tulip (2005) has documented that the Greenbook forecast errors for output
were largest early in the sample period.
To assess the importance of the early sample in our results, we construct results
analogous to Table 3 for the period since 1984 (Table 5). As expected, the RMSPEs
are substantially smaller than for the whole period in almost every case. Our qualitative
conclusions, however, remain essentially the same. One conclusion is somewhat strength-
ened by excluding the turbulent period of the early 1980s: for forecasting growth, the
iterated AR now performs even better compared to large dataset methods than before.
5.2 Bias versus standard deviation
In Table 6 we show the decomposition of the RMSPEs for the Greenbook forecasts and
16
statistical forecasts using data up to quarter t 1 into bias and standard deviation com-
ponents. Results are shown for both the full sample and the period since 1984. For the
in‡ation forecasts, all have some upward bias, at least at horizons beyond two quarters.
The bias is worse for the stationary univariate and factor model time series forecasts
than it is for the Greenbook forecasts, the random walk, IFAV and UCSV forecasts, and
the predictions based on Bayesian model averaging. The standard deviations of the in-
‡ation forecasts can be thought of as bias-adjusted versions of the RMSPEs. As can be
seen in Table 6, the standard deviations of the in‡ation forecasts are quite similar to the
RMSPEs that we reported in Tables 3 and 5. The Greenbook in‡ation forecast errors
have a smaller standard deviation than any of the statistical forecasts. Turning to the
growth forecasts, the Greenbook forecasts have a modest downward bias, while all of the
statistical models have an upward bias, except for the random walk and UCSV forecasts.
The standard deviations of the growth forecasts are again very similar to the RMSPEs
that we reported in Tables 3 and 5.
5.3 Comparison with Atkeson and Ohanian
In contrast to our results, Atkeson and Ohanian (2001) compared the RMSPE of Green-
book and random walk forecasts of in‡ation, and found that they were roughly equal,
concluding that there was no incremental information in the Greenbook. The conclusion
was based on 13 observations taken from the last Greenbook forecast in each year from
1983 to 1995 inclusive. They compared the Greenbook projection of GNP/GDP de‡ator
in‡ation over the subsequent four quarters with a random walk forecast computed as the
value of in‡ation over the previous four quarters. Thus, their measure of in‡ation and
their version of the random walk forecast are di¤erent from ours.
We came very close to replicating their results by using their de…nitions and limiting
our focus to the 13 observations they studied. We …nd that the RMSPE of the Greenbook
forecast, relative to that of the random walk, was 0.96. Using their de…nitions but our
17
full sample, the ratio is more favorable to Greenbook at 0.73.
We examine the sampling properties of the estimate based on 13 observations using
the same bootstrap procedure used above. Figure 1 shows the bootstrap approximation
to the sampling distribution of the ratio of RMSPEs. It can clearly be seen that while
this is centered around 1, in line with the Atkeson and Ohanian result, the ratio is quite
imprecisely estimated. A 95 percent con…dence interval for this ratio would span from
0.67 to 1.28, which includes the estimates for this ratio that we obtain with our much
larger sample size. Thus, the good performance of the random walk 13 observations is
entirely consistent with our …nding that Greenbook dominates the random walk forecast.
5.4 Real-time versus ex-post revised data
All of the results presented so far use our real-time dataset. Most earlier work in the
literature is based on a single recent vintage of data, and, hence, is not strictly comparable
to our work. For comparison purposes, and to shed light on the importance of vintage
issues, we repeat our exercise using a single vintage of data as observed in the Greenbook
dated December 14, 2000 (the last vintage that is outside of the …ve-year window during
which the data are con…dential).
In this experiment, for each Greenbook, we took the Greenbook forecasts and com-
pared these with the data as observed in the December 2000 vintage data. Then we
constructed each of the statistical forecasts described above using observations for all
the predictors from the December 2000 Greenbook at the quarterly frequency using a
recursive out-of-sample scheme. For example, we used the December 2000 Greenbook
vintage data from 1960Q2 through 1995Q3 to construct forecasts for each quarter from
1995Q4 through 1996Q4 (horizons 0 through 5). And we again compared these fore-
casts with the data as observed in the December 2000 Greenbook. Only forecasts for
2000Q3 and prior quarters are included, because data for subsequent quarters were not
yet available in December 2000, and so would not be used by a researcher in a standard
18
forecast evaluation exercise.
The RMSPEs are shown in Table 7. The RMSPEs obtained using ex-post revised
data are uniformly smaller than their counterparts using real-time data (Table 3). Also,
the atheoretical forecasts perhaps fare a little better relative to Greenbook in the revised
dataset than using real-time data. This is true for both the univariate forecasts and
for the forecasts using a large number of predictors. The relative performance of the
large-dataset methods and univariate forecasts is however comparable in the real-time
and revised datasets.
Our dataset is large, but not as large as some datasets using many predictors, such
as the datasets used in Stock and Watson (2002, 2005). Bernanke and Boivin (2003)
found that factor models gave better forecasts when using the 215-variable database of
Stock and Watson (2002) than when using their real-time database of 78 variables, or
when using the ex-post revised observations on these 78 variables.
For comparison purposes, we re-did the analysis of the previous section but using the
database of Stock and Watson (2005) which is, of course, revised data.
12
The RMSPEs
are shown in Table 8. The FAAR, factor-augmented VAR and dynamic factor forecasts
give considerably better forecasts of in‡ation when using this dataset than when using
the …nal (December 2000) vintage of our Greenbook dataset. The same is true, to a
much lesser extent, for the IFAV forecasts that impose a unit root. The results in Tables
7 and 8 are otherwise not very di¤erent, but, for in‡ation, factor forecasts seem to work
best when using more highly disaggregated data than we have.
These results present the same puzzle as noted by Bernanke and Boivin (2003): the
factor models perform much more poorly in forecasting in‡ation in fairly large real-time
datasets than in the much the larger, fully-revised dataset of Stock and Watson. There
are three natural hypotheses about this puzzle. First, it may be that the data revisions
12
This dataset contains 135 variables and is larger than our database, but still smaller than the 215-
variable database of Stock and Watson (2002) that does not however extend through 2000.
19
account for the di¤erence. Second, it may be that the larger dataset simply has more
relevant information than the smaller one. Third, it may be that while the forecast-
relevant information may be similar in the large and small datasets, the factor models
extract that information more e¤ectively in the larger dataset.
With the available data, we cannot fully discriminate among these three hypotheses.
The results of Table 7 and analogous results of Bernanke and Boivin suggest that the
…rst hypothesis is not the main story since, with our datasets, the relative performance
of the factor-based forecasts and the univariate time series forecasts is comparable in the
real-time data (Table 3) and the revised data (Table 7).
Our results also suggest that the second hypothesis may not be a major factor.
Despite the fact that the performance of the factor models varies considerably as we
consider di¤erent datasets, the model averaging approaches give good performance and
similar performance throughout. For example, the Table 7 and Table 8 results for the
model averaging forecast are very similar. Thus, there is no strong support for the view
that the larger dataset contains important information about output and in‡ation that
is simply missing from the smaller dataset.
Turning to the third hypothesis, it may be that adding more variables—or adding
the particular mix of variables that happen to be added by Stock and Watson (2002,
2005)—allows the factor methods more clearly to isolate the important information that
the model averaging methods more robustly isolate across the various cases. Perhaps this
should not be surprising. The factor model approaches merely pick out a few dominant
factors from the large dataset, not focusing on the ability of those factors to forecast
in‡ation or output directly. Loosely speaking, we never ask the full data the question we
are ultimately interested in. One might suspect that such methods could be sensitive to
which variables happen to be included.
In contrast, the model averaging methods directly exploit the information about
each variable for the question of interest. But these methods do so in a very limited
20
way—we consider a large number of bivariate models. The the distinction between these
methods is in part a distinction between bivariate and more richly multivariate methods.
Of course, the relative merits of these methods could be formalized and assessed as
hypotheses about the underlying factor structure in macro data. We leave this analysis
for future work.
5.6 Combining atheoretical and Greenbook forecasts
The atheoretical and Greenbook forecasts have comparable predictive performance for
growth beyond the current quarter, though not for in‡ation. Meanwhile, since they are
not perfectly correlated, this implies that a better forecast must be available by combining
the Greenbook and atheoretical forecasts, at least for growth. Letting ^ n
GB
T+h
and ^ n
BMA
T+h
denote the Greenbook and BMA forecasts, respectively, we can de…ne a combined forecast
as
^ n
COMB
T+h
= \^ n
GB
T+h
+ (1 \)^ n
BMA
T+h
For any …xed \, we can then evaluate the real-time RMSPE of this forecast and plot
it against \. These are shown in Figures 2 and 3 for in‡ation and growth forecasts
respectively. The jumping o¤ point is t 1, and results for three horizons are shown.
For in‡ation, the optimal choice of \ is close to 1, which amounts to putting all of the
weight on the Greenbook forecast. For growth, however, the optimal choice of \ is about
0.5 at horizons beyond the current quarter, meaning that substantial shrinkage of the
Greenbook towards the BMA forecast (or vice-versa) would give better predictions than
either forecast alone. The reductions in RMSPE due to this shrinkage are, however, very
modest.
5.6 The Greenbook as a conditional forecast
The Greenbook projection is conditioned on a hypothetical path of policy that is coun-
terfactual in the sense that it is not supposed to be a forecast of policy (see Faust and
21
Wright (2006)). Nearly all work on assessing the information content of central bank
forecasts ignores their conditional nature and assesses the forecasts as though they were
unconditional forecasts. We have done so here, thereby implicitly assessing their prop-
erties when viewed as though they were unconditional forecasts. However it should be
noted that Greenbook could appear better—or worse—if we were to take proper account
of its conditional nature. Faust and Wright propose a method for backing out an im-
plied unconditional forecast based on comparing the hypothetical path of policy with
the path implied by money market futures rates, under the assumption that the latter
is the unconditional expectation of future interest rates. In future work, we intend to
include this implied unconditional counterpart of the Greenbook forecast in the forecast
evaluation exercise.
6. Conclusions
In this paper, we compare the forecast accuracy of three types of forecasting models: small
time series models that are atheoretical from an economic perspective, large atheoretical
time series models, and the Federal Reserve’s Greenbook forecast. We focus on predicting
in‡ation and output growth in a large real-time dataset.
The Greenbook forecast comparison is interesting for a few reasons. This forecast
is generally thought to be at or near the frontier of best accuracy among real-world
forecasts. Further, it is based on a much wider array of information than even our large
time-series models, and it is subjective and heavily in‡uenced by the sta¤’s understanding
of economics. Thus, to some extent, this comparison allows us to measure whether one
version of a much richer forecasting framework helps much in forecasting.
Sims (2002) and others have conjectured that this richer framework of the Green-
book forecast may be of most value in evaluating the current state of the economy—one
can take account of strikes, hurricanes, and the timing of holidays. Also some of the
ability of the Greenbook to assess the current state of the economy surely owes to the
22
attention that Federal Reserve sta¤ pay to incoming macroeconomic data releases. The
case is particularly clear for CPI in‡ation as these data are released at the monthly
frequency. Depending on the time within the quarter when the Greenbook forecast is
made, one or even two months of CPI in‡ation for the current quarter will already have
been released and will clearly be factored into the current-quarter CPI in‡ation forecast.
Authors including Giannone, Reichlin and Small (2007) and Aruoba, Diebold and Scotti
(2007) have recently proposed powerful methods for measuring the current state of the
economy from the real-time monitoring of incoming data releases.
One naturally wonders whether the advantage of Greenbook relative to the atheo-
retical methods conisdered in this paper decays quickly as the forecasting horizon grows.
We examine this question by checking whether one could improve forecast accuracy by
using hybrid forecast that is the Greenbook forecast through some horizon such as 1 or
2 quarters out and then switching to a time series method.
For both output and in‡ation, we …nd that the Greenbook forecast is quite good
compared with purely atheoretical methods. Beyond this, the results for output and
in‡ation are strikingly di¤erent.
For output, once we give the time series models the Greenbook forecast for the
current quarter at the time of the forecast, the atheoretical models perform as well or
better than Greenbook. Indeed, after the current quarter nothing does much better than
the AR(1) model. Given a rich evaluation of the current state of the economy, it is hard
to beat nearly the simplest of time series models.
For in‡ation, Greenbook seems to have a considerable advantage over the atheo-
retical methods even at long horizons and even when the atheoretical methods jump o¤
only after several quarters of Greenbook forecast. We …nd it surprising that Greenbook
continues to dominate, even at long horizons and that the results for output and in‡ation
are so di¤erent.
The results also shed some light on comparisons among the competing time series
23
methods. Two summary points are worth emphasizing. First, the results are generally
less supportive of the large model methods than some earlier work. The large models
o¤er no substantial gain for output growth prediction and only the best of the methods
o¤er an advantage for in‡ation. Second, the large model forecasts that are built up using
an average of bivariate models consistently outperform those that are more richly mul-
tivariate. That is, Bayesian and simple model averaging tend to dominate the models
attempting to exploit some richer factor structure in the macro data. This result is con-
sistent with earlier work using a single vintage and with earlier work on model averaging
more generally. The results raise potentially important questions about the underlying
factor structure of the macro data.
References
Ang, A., G. Bekaert and M. Wei (2006): Do Macro Variables, Asset Markets or Surveys
Forecast In‡ation Better?, Journal of Monetary Economics, forthcoming.
Aruoba, S.B., F.X. Diebold and C. Scotti (2007): Real-Time Measurement of Business
Conditions, working paper.
Atkeson, A. and L.E. Ohanian (2001): Are Phillips Curves Useful for Forecasting In‡a-
tion?, Federal Reserve Bank of Minneapolis Quarterly Review, 25, pp.2-11.
Bates, J.M. and C.W.J. Granger (1969): The Combination of Forecasts, Operations
Research Quarterly, 20, pp.451-468.
Bernanke, B.S. and J. Boivin (2003): Monetary Policy in a Data-Rich Environment,
Journal of Monetary Economics, 50, pp.525-546.
Bernanke, B.S., J. Boivin and P. Eliasz (2005): Measuring Monetary Policy: A Fac-
tor Augmented Vector Autoregressive (FAVAR) Approach, Quarterly Journal of
Economics, 120, pp.387-422.
24
Boivin, J. and S. Ng (2005): Monetary Policy in a Data-Rich Environment, Journal of
Monetary Economics, 50, pp.525-546.
Clark, T.E. and M. McCracken (2001): Tests of Equal Forecast Accuracy and Encom-
passing for Nested Models, Journal of Econometrics, 105, pp.85-110.
D’Agostino, A. and D. Giannone (2006): Comparing Alternative Predictors Based on
Large Panel Factor Models, European Central Bank Working Paper 680.
Diebold, F.X. and R.S. Mariano (1995): Comparing Predictive Accuracy, Journal of
Business and Economic Statistics, 13, pp.253-263.
Edge, R.M., M. T. Kiley and J.P. Laforte (2006): A Comparison of Forecast Performance
between Federal Reserve Sta¤ Forecasts, Simple Reduced form Models and a DSGE
Model, working paper.
Faust, Jon and Jonathan H. Wright (2006): E¢cient Forecast Tests for Conditional
Policy Forecasts, working paper.
Fernandez, C., E. Ley and M.F.J. Steel (2001): Model Uncertainty in Cross-Country
Growth Regressions, Journal of Applied Econometrics, 16, pp.563-576.
Forni, M., M. Hallin, M. Lippi and L. Reichlin (2005): The Generalized Dynamic
Factor Model: One-Sided Estimation and Forecasting, Journal of the American
Statistical Association, 100, pp.830-840.
Giannone, D., L. Reichlin and L. Sala (2004): Monetary Policy in Real Time, in M.
Gertler and K. Rogo¤ (eds.), NBER Macroeconomics Annual, 2004, MIT Press,
Cambridge.
Giannone, D., L. Reichlin and D. Small (2007): Nowcasting GDP: The Real Time
Informational Content of Macroeconomic Data Releases, working paper.
25
Koop, G. and S. Potter (2003): Forecasting in Large Macroeconomic Panels Using
Bayesian Model Averaging, Federal Reserve Bank of New York Sta¤ Report 163.
Künsch (1989): The Jackknife and the Bootstrap for General Stationary Observations,
Annals of Statistics, 17, pp.1217-1241.
Liu, R. and K. Singh (1992): Moving Blocks Jackknife and Bootstrap Capture Weak
Dependence, in Exploring the Limits of the Bootstrap (R. Lepage and L. Billard,
eds.), Wiley, New York.
Marcellino, M., J.H. Stock and M.W. Watson (2006): A Comparison of Direct and It-
erated Multistep AR methods for Forecasting Macroeconomic Time Series, Journal
of Econometrics, 135, pp.499-526.
Reifschneider, D.L., D.J. Stockton and D.W. Wilcox (1997): Econometric Models and
the Monetary Policy Process, Carnegie Rochester Series on Public Policy, 47, pp.1-
37.
Romer, C.D. and D.H. Romer (2000): Federal Reserve Information and the Behavior of
Interest Rates, American Economic Review, 90, pp.429-457.
Sims, C.A. (2002): The Role of Models and Probabilities in the Monetary Policy
Process, Brookings Papers on Economic Activity, 2, pp.1-40.
Stock, J.H. and M.W. Watson (1999): Forecasting In‡ation, Journal of Monetary
Economics, 44, pp.293-335.
Stock, J.H. and M.W. Watson (2002): Forecasting Using Principal Components from
a Large Number of Predictors, Journal of the American Statistical Association, 97,
pp.1167-1179.
26
Stock, J.H. and M.W. Watson (2003): Forecasting Output and In‡ation: The Role of
Asset Prices, Journal of Economic Literature, 41, pp.788-829.
Stock, J.H. and M.W. Watson (2005): An Empirical Comparison of Methods for
Forecasting Using Many Predictors, working paper.
Stock, J.H. and M.W. Watson (2007): Has In‡ation Become Harder to Forecast?,
Journal of Money, Credit and Banking, 39, pp.3-34.
Tetlow, R. and B. Ironside (2004): Real-Time Model Uncertainty in the United States:
the Fed from 1996 to 2003, working paper.
Tulip, P. (2005b): Has Output Become More Predictable? Changes in Greenbook Fore-
cast Accuracy, Finance and Economics Discussion Series 2005-31, Federal Reserve
Board.
Wright, J.H. (2003): Forecasting U.S. In‡ation by Bayesian Model Averaging, Interna-
tional Finance Discussion Paper 780, Federal Reserve Board.
27
Table 1: Dates of Greenbook Forecasts Employed
Date Date Date Date
19800314 19850515 19901107 19961106
19800416 19850703 19901212 19961212
19800514 19850814 19910130 19970129
19800702 19850925 19910320 19970319
19800806 19851030 19910508 19970515
19800910 19851211 19910626 19970625
19801015 19860205 19910814 19970814
19801112 19860326 19910925 19970924
19801212 19860514 19911030 19971106
19810128 19860702 19920130 19971211
19810325 19860813 19920325 19980128
19810513 19860917 19920514 19980325
19810701 19861029 19920624 19980514
19810812 19861210 19920813 19980624
19810930 19870204 19920930 19980817
19811110 19870325 19921112 19980923
19811216 19870513 19921216 19981112
19820324 19870701 19930129 19981216
19820512 19870812 19930317 19990128
19820623 19870916 19930514 19990324
19820929 19871028 19930630 19990513
19821110 19871209 19930811 19990625
19821215 19880203 19930915 19991001
19830202 19880323 19931110 19991215
19830323 19880511 19931215 20000127
19830518 19880622 19940131 20000327
19830706 19881207 19940316 20000512
19830817 19890201 19940513 20000621
19830928 19890322 19940629 20000816
19831109 19890510 19940921 20000927
19831214 19890628 19941109 20001108
19840125 19890816 19941214 20001214
19840321 19890927 19950125
19840516 19891108 19950322
19840711 19891213 19950517
19840815 19900131 19950628
19840926 19900321 19950920
19841031 19900509 19951108
19841212 19900627 19951214
19850206 19900815 19960626
19850320 19900926 19960918
This table lists the dates of publication of each Greenbook in our sample.
28
Table 2. Variables, transforms, and coverage
Variable Transform First Last Number
Consumer spending on durable goods, real DLN 19800314 20001214 145
Consumer spending on durable goods, nom DLN 19800314 19951214 130
Exports, real DLN 19800314 20001214 145
Exports, nom DLN 19800314 20001214 145
Government spending, real DLN 19960626 20001214 34
Government spending, nominal DLN 19800314 19951214 130
Residential construction spending, real DLN 19800314 20001214 145
Residential construction spending, nom DLN 19800314 20001214 140
Inventory investment, nominal FD 19800314 20001214 145
Inventory investment, real FD 19800314 20001214 145
Inventory investment, manuf. and trade, real FD 19960626 19980624 16
Imports, real DLN 19800314 20001214 145
Imports, nom DLN 19800314 20001214 145
Pesronal consumption spending, real DLN 19800314 20001214 145
Pesronal consumption spending, nom DLN 19800314 20001214 145
Spending on producers’ durable equipment, real DLN 19800314 20001214 145
Spending on producers’ durable equipment, nom DLN 19800314 20001214 140
Spending on producers’ structures, real DLN 19800314 20001214 145
Spending on producers’ structures, nom DLN 19800314 20001214 140
Merchandise exports DLN 19800314 19920514 103
Foreign nom GNP index (bilateral weights) DLN 19821215 19911030 69
Foreign short-term interest rate Level 19800314 19930129 109
Stock of consumer durables, real DLN 19800314 20001214 145
Stock of nonfarm inventories, real DLN 19800314 20001214 148
Stock of nonfarm inventories, ex manuf. and trade DLN 19960626 19980624 16
Stock of nonfarm nondurable inventories, real DLN 19800314 19880622 76
Stock of nonfarm nonretail inventories, real DLN 19800314 19951214 130
Stock of nonfarm retail durable inventories, real DLN 19800314 19951214 130
Net stock of producers’ durable equipment, real DLN 19800314 20001214 145
Net stock of producers’ structures, real DLN 19800314 20001214 145
29
Table 2. Variables, transforms, and coverage, cont.
Variable Transform First Last Number
Nonfarm business employment DLN 19800314 20001214 145
Employment of nonfarm proprietors DLN 19800314 19951214 130
Hours of employees; nonfarm business sector DLN 19800314 20001214 145
Unemployment Level 19800314 19951214 130
M1 DLN 19800314 19951214 109
M2 DLN 19800314 20001214 157
Commercial and industrial loans at banks DLN 19800314 19921216 108
Nonborrowed reserves DLN 19800314 19951214 130
Currency plus travelers checks DLN 19800314 19951214 130
Durable goods consumption de‡ator DLN 19800314 19951214 130
Consumption de‡ator DLN 19800314 19951214 130
GNP de‡ator DLN 19800314 19951214 130
Average price per barrel of imported oil DLN 19800314 20001214 145
Wholesale price index for fuels DLN 19800314 19951214 130
Employee compensation per hour DLN 19800314 20001214 145
Foreign exchange rate index, bilateral weights DLN 19800314 19911030 100
Corporate bond rate; Moody’s seasoned AAA Level 19800314 20001214 145
Dividend-price ratio (Standard and Poors) Log 19800314 20001214 145
Federal funds rate Level 19800314 20001214 145
Ten-year Treasury CM yield Level 19800314 20001214 145
3 month Treasury bill rate Level 19800314 20001214 145
Real GNP DLN 19800314 19951214 130
Nominal GNP DLN 19800314 19951214 130
Merchandise trade balance FD 19800314 19921112 107
Net exports FD 19800314 19951214 130
Compensation of employees DLN 19800314 19951214 130
Corporate pro…ts DLN 19800314 20001214 145
Foreign output DLN 19790516 20001214 163
Foreign consumer price index DLN 19790516 19930129 108
Foreign exchange rate, multilateral weights DLN 19790516 19930129 108
Hours; household and institutions sector DLN 19791114 19921112 98
30
Table 2. Variables, transforms, and coverage, cont.
Variable Transform First Last Number
Total reserves DLN 19791114 19921216 102
Civilian unemployment rate Level 19801112 19951214 92
Stock of autos DLN 19820512 19890816 55
Stock of consumer durables ex autos, real DLN 19820512 19951214 103
Corporate bond rate Level 19820512 20001214 137
Mortgage rate; e¤ective annual yield Level 19820623 20001214 136
Foreign real GNP index; bilateral weights DLN 19800314 19911030 100
Output per hour DLN 19821215 19951214 99
Final sales, real DLN 19840321 20001214 107
Final sales, nominal DLN 19991001 20001214 10
Spending on business …xed investment, real DLN 19841212 19951214 83
Spending on business …xed investment, nom DLN 19841212 19951214 83
Nominal exchange rate DLN 19850320 19930129 60
Total cilivian employment DLN 19870204 19951214 66
Stock of nonfarm nondurable inventories, real DLN 19881207 19951214 54
Wholesale price of Petroleum products DLN 19890510 19951214 51
Stock of motor vehicles and parts, real DLN 19890816 19951214 49
Commodity price, industrial materials DLN 19900321 19901107 2
Consumer Sentiment Level 19920130 19921112 7
GDP Implicit De‡ator DLN 19920130 20001214 64
ReaL GDP DLN 19920130 20001214 64
Nominal GDP DLN 19920130 20001214 64
Current Account Balance FD 19920130 20001214 64
Total Housing Stock DLN 19920514 20001214 62
CPI DLN 19920514 19951214 28
CPI ex food and energy DLN 19920514 19951214 28
Producer Price Index DLN 19920624 19951214 27
Real Disposable Income DLN 19920624 19951214 27
Nominal Disposable Income DLN 19921216 19951214 23
G10 real exchange rate DLN 19930915 20001214 51
G18 real exchange rate DLN 19930317 19930811 4
31
Table 2. Variables, transforms, and coverage, cont.
Variable Transform First Last Number
G18 nom exchange rate DLN 19930317 19930811 4
Civilian employment DLN 19940316 19951214 13
Civilian labor force DLN 19940316 20001214 47
Price de‡ator for crude energy consumption DLN 19950517 20001214 39
Compensation per hour DLN 19950517 19951214 5
Consumption, energy sector DLN 19960626 20001214 34
Import Price of Petroleum products DLN 19960626 20001214 34
E¤ective federal funds rate Level 19960626 20001214 34
E¤ective ten-year TreasuryCM yield Level 19960626 20001214 34
E¤ective three-month Treasury bill yield Level 19960626 20001214 34
Core PCE price index DLN 19970814 20001214 26
Labor Force Participation Rate Level 19980817 20001214 18
Moody’s BAA corporate bond yield Level 19990513 20001214 12
Final sales of GDP DLN 19991001 20001214 10
Gross private domestic investment DLN 20000621 20001214 5
Nonfarm business sector workweek LN 20000621 20001214 5
Personal saving DLN 20000621 20001214 5
Notes: For transforms, DLN means …rst di¤erence of the natural logarithm, FD means …rst di¤erence, Log means natural
logarithm, level means no transform. The columns labelled …rst and last give the date of the …rst and last vintage in which
the variable is present, respectively. The column labelled num gives the number of vintages in which the variable is present.
32
Table 3a: RMSPE of de‡ator in‡ation forecasts, 1979-2000
hor. GB RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
jump o¤ -1
0 0.81 1.31 1.09 1.09 1.07 1.06 1.18 1.28 1.24 1.30 1.12
1 0.84 1.50 1.26 1.26 1.22 1.17 1.23 1.58 1.42 1.63 1.24
2 0.91 1.27 1.16 1.17 1.13 1.11 1.22 1.65 1.17 1.61 1.18
3 1.03 1.29 1.32 1.34 1.31 1.10 1.48 1.94 1.18 1.78 1.29
4 1.06 1.40 1.45 1.47 1.47 1.19 1.65 2.21 1.42 1.90 1.39
5 1.30 1.71 1.72 1.78 1.75 1.39 1.84 2.43 1.65 2.15 1.66
jump o¤ 0
1 0.84 1.16 1.16 1.16 1.15 1.13 1.26 1.40 1.22 1.39 1.14
2 0.91 1.26 1.14 1.13 1.11 1.08 1.21 1.64 1.17 1.45 1.18
3 1.03 1.27 1.23 1.23 1.20 1.06 1.36 1.84 1.08 1.53 1.24
4 1.06 1.41 1.41 1.43 1.43 1.15 1.60 2.09 1.27 1.78 1.39
5 1.30 1.65 1.69 1.72 1.71 1.37 1.72 2.43 1.54 2.13 1.64
jump o¤ 1
2 0.91 1.19 1.08 1.08 1.08 1.08 1.18 1.40 1.13 1.32 1.15
3 1.03 1.43 1.26 1.26 1.24 1.05 1.30 1.68 1.28 1.45 1.31
4 1.06 1.41 1.34 1.35 1.34 1.11 1.44 1.87 1.26 1.58 1.36
5 1.30 1.58 1.58 1.62 1.61 1.28 1.71 2.22 1.43 1.90 1.57
jump o¤ 2
3 1.03 1.16 1.21 1.21 1.21 1.11 1.34 1.49 1.20 1.36 1.21
4 1.06 1.27 1.31 1.33 1.31 1.11 1.40 1.70 1.19 1.40 1.29
5 1.30 1.44 1.53 1.54 1.53 1.26 1.63 2.08 1.31 1.68 1.48
jump o¤ 3
4 1.06 1.18 1.23 1.23 1.23 1.08 1.34 1.47 1.14 1.32 1.22
5 1.30 1.44 1.48 1.47 1.46 1.23 1.57 1.90 1.32 1.56 1.45
Notes: Each column, 2–11, reports the root mean square prediction error (RMSPE) for the
forecast labelled at top for the current quarter (hor= 0) through 5 quarters into the future.
For forecasts other than Greenbook, the data on which the forecast is based is the real-
time data augmented to bring the data up to the jumping o¤ point using the Greenbook
forecast. Bold indicates that the alternative forecast has smaller RMSPE than Greenbook;
,, indicate that the di¤erence between the RMSPE for the model and Greenbook is
statistically signi…cant at the 1, 5, or 10 percent level, respectively, based on the bootstrap
p-values for the DM test as described in the text.
33
Table 3b: RMSPE of CPI in‡ation forecasts, 1979-2000
hor. GB RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
jump o¤ -1
0 0.82 2.05 1.75 1.75 1.72 1.62 1.66 1.85 1.93 1.76 1.92
1 1.66 2.20 1.80 1.82 1.78 1.56 1.89 2.15 2.06 1.94 1.92
2 1.78 1.95 1.85 1.95 1.91 1.58 1.95 2.11 1.73 1.94 1.93
3 1.83 2.51 2.35 2.37 2.31 1.86 2.28 2.61 2.26 2.12 2.42
4 1.60 2.57 2.23 2.22 2.19 1.85 2.50 2.77 2.51 2.29 2.36
5 1.63 2.44 2.25 2.20 2.12 2.21 2.43 2.70 2.35 2.19 2.38
jump o¤ 0
1 1.66 2.19 1.80 1.80 1.78 1.65 1.75 2.04 2.11 1.73 1.93
2 1.78 2.23 1.82 1.82 1.79 1.55 1.87 2.24 2.02 1.71 1.96
3 1.83 2.13 2.01 2.04 2.00 1.63 2.01 2.38 1.95 1.85 2.06
4 1.60 2.52 2.32 2.41 2.37 1.90 2.41 2.65 2.45 2.17 2.33
5 1.63 2.61 2.23 2.26 2.19 1.71 2.33 2.86 2.47 2.17 2.40
jump o¤ 1
2 1.78 1.98 1.79 1.79 1.78 1.60 1.77 2.01 1.92 1.79 1.89
3 1.83 2.11 2.05 2.08 2.04 1.68 1.99 2.27 1.90 1.86 2.04
4 1.60 2.21 2.31 2.46 2.43 1.97 2.38 2.34 2.10 1.90 2.18
5 1.63 2.40 2.27 2.30 2.27 1.74 2.22 2.64 2.26 2.06 2.35
jump o¤ 2
3 1.83 1.94 1.92 1.92 1.90 1.70 1.86 2.10 1.90 1.93 1.93
4 1.60 1.86 2.07 2.13 2.10 1.76 2.08 2.14 1.81 1.81 1.92
5 1.63 2.16 2.24 2.33 2.30 1.85 2.18 2.48 2.12 1.84 2.18
jump o¤ 3
4 1.60 1.70 1.77 1.77 1.75 1.62 1.76 1.86 1.66 1.77 1.74
5 1.63 1.88 2.02 2.06 2.02 1.70 2.09 2.18 1.86 1.80 1.95
Notes: Each column, 2–11, reports the root mean square prediction error (RMSPE) for the
forecast labelled at top for the current quarter (hor= 0) through 5 quarters into the future.
For forecasts other than Greenbook, the data on which the forecast is based is the real-
time data augmented to bring the data up to the jumping o¤ point using the Greenbook
forecast. Bold indicates that the alternative forecast has smaller RMSPE than Greenbook;
,, indicate that the di¤erence between the RMSPE for the model and Greenbook is
statistically signi…cant at the 1, 5, or 10 percent level, respectively, based on the bootstrap
p-values for the DM test as described in the text.
34
Table 3c: RMSPE of output growth forecasts, 1979-2000
hor. GB RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
jump o¤ -1
0 2.17 3.31 2.77 2.77 2.71 2.73 2.75 2.73 2.73 2.75 2.74
1 2.75 3.56 2.75 2.76 2.64 2.49 2.80 2.63 2.81 2.67 2.72
2 2.72 3.82 2.83 2.90 2.80 2.86 2.88 2.69 2.97 2.92 2.96
3 2.76 3.40 2.61 2.58 2.54 2.66 2.83 2.74 3.09 2.63 2.84
4 2.57 4.08 2.69 2.71 2.62 2.59 2.56 2.60 3.11 2.67 2.97
5 2.41 3.62 2.66 2.74 2.71 2.86 2.58 2.60 3.62 2.76 2.92
jump o¤ 0
1 2.75 2.90 2.66 2.66 2.57 2.48 2.59 2.51 2.76 2.50 2.61
2 2.72 3.20 2.76 2.82 2.71 2.53 2.74 2.58 2.79 2.86 2.86
3 2.76 3.08 2.61 2.63 2.58 2.69 2.76 2.69 2.73 2.66 2.72
4 2.57 3.37 2.66 2.66 2.58 2.58 2.52 2.57 2.94 2.65 2.89
5 2.41 3.19 2.64 2.74 2.65 2.61 2.54 2.53 2.91 2.83 2.79
jump o¤ 1
2 2.72 3.07 2.74 2.74 2.68 2.68 2.66 2.61 2.63 2.72 2.79
3 2.76 3.04 2.62 2.61 2.58 2.64 2.75 2.64 2.82 2.69 2.75
4 2.57 2.83 2.57 2.57 2.51 2.50 2.58 2.60 2.68 2.63 2.71
5 2.41 2.74 2.62 2.62 2.56 2.56 2.57 2.53 2.79 2.71 2.62
jump o¤ 2
3 2.76 2.91 2.71 2.71 2.66 2.58 2.62 2.56 2.80 2.60 2.77
4 2.57 2.62 2.54 2.51 2.46 2.44 2.53 2.55 2.64 2.57 2.62
5 2.41 2.48 2.59 2.66 2.59 2.50 2.61 2.52 2.73 2.65 2.53
jump o¤ 3
4 2.57 2.68 2.55 2.55 2.49 2.43 2.44 2.44 2.59 2.52 2.61
5 2.41 2.54 2.54 2.59 2.52 2.33 2.52 2.44 2.66 2.52 2.47
Notes: Each column, 2–11, reports the root mean square prediction error (RMSPE) for the
forecast labelled at top for the current quarter (hor= 0) through 5 quarters into the future.
For forecasts other than Greenbook, the data on which the forecast is based is the real-
time data augmented to bring the data up to the jumping o¤ point using the Greenbook
forecast. Bold indicates that the alternative forecast has smaller RMSPE than Greenbook;
,, indicate that the di¤erence between the RMSPE for the model and Greenbook is
statistically signi…cant at the 1, 5, or 10 percent level, respectively, based on the bootstrap
p-values for the DM test as described in the text.
35
Table 4a: Percentage of time series de‡ator in‡ation forecasts that are more
accurate than GB, 1979-2000
hor. RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
jump o¤ -1
0 39 34 34 30 32 30 27 37 33 35
1 27 25 29 32 32 33 27 30 21 34
2 41 36 34 38 37 41 17 39 26 37
3 42 37 33 35 39 37 22 47 22 34
4 35 27 26 26 37 28 14 40 17 41
5 38 23 19 21 47 30 16 37 19 41
jump o¤ 0
1 32 31 31 31 32 30 33 31 28 32
2 34 34 34 33 34 38 23 35 22 37
3 42 38 34 36 39 32 26 44 24 39
4 34 29 30 27 39 34 18 42 23 40
5 35 28 26 23 43 32 16 41 18 45
jump o¤ 1
2 38 36 36 37 34 40 32 41 22 39
3 35 33 32 34 39 33 24 37 25 40
4 36 31 31 30 37 31 21 37 22 39
5 36 36 32 32 44 31 19 39 23 36
jump o¤ 2
3 40 36 36 38 38 39 26 39 26 39
4 32 32 32 32 36 32 28 34 27 30
5 37 31 31 29 37 30 23 39 24 41
jump o¤ 3
4 39 34 34 35 39 33 28 39 30 37
5 41 33 32 31 40 30 24 42 28 37
Notes: Each column, 2–10, reports the share of forecast periods in which the GB forecast
error is larger, ex post, in absolute value than the forecast error from the model labelled at
the top. Entries greater than 50 percent indicate that the model forecast does better than
GB in more than half of forecast periods and are in bold. See also notes to Table 3.
36
Table 4b: Percentage of time series CPI in‡ation forecasts that are more
accurate than GB, 1979-2000
hor. RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
jump o¤ -1
0 15 14 14 16 22 17 16 13 17 17
1 42 45 45 45 49 40 35 43 41 43
2 44 43 42 43 46 43 39 50 36 45
3 41 37 33 34 52 33 33 31 34 37
4 42 36 37 39 54 35 28 39 28 43
5 44 30 33 33 55 38 28 36 23 42
jump o¤ 0
1 37 39 39 39 44 44 39 38 40 48
2 45 51 50 52 56 46 37 46 43 52
3 48 47 45 46 56 43 43 48 46 48
4 37 37 34 34 54 39 25 36 30 33
5 44 34 32 35 61 37 31 43 26 37
jump o¤ 1
2 42 48 48 48 61 52 40 50 45 49
3 46 47 45 48 58 45 39 48 44 50
4 49 39 38 41 53 46 34 48 38 45
5 39 32 30 29 54 39 23 42 28 33
jump o¤ 2
3 46 51 51 53 63 54 42 55 42 52
4 52 41 41 41 59 46 40 51 43 49
5 47 32 34 38 59 44 31 46 41 39
jump o¤ 3
4 50 48 48 51 53 50 43 55 48 46
5 43 38 36 41 59 46 37 50 47 38
Notes: Each column, 2–10, reports the share of forecast periods in which the GB forecast
error is larger, ex post, in absolute value than the forecast error from the model labelled at
the top. Entries greater than 50 percent indicate that the model forecast does better than
GB in more than half of forecast periods and are in bold. See also notes to Table 3.
37
Table 4c: Percentage of time series output growth forecasts that are more
accurate than GB, 1979-2000
hor. RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
jump o¤ -1
0 25 37 37 39 41 38 34 36 37 31
1 39 54 57 51 48 41 43 43 46 46
2 39 51 52 54 52 45 47 45 50 39
3 45 54 55 53 50 47 43 41 52 50
4 36 50 44 45 45 40 43 41 44 40
5 36 41 43 46 40 43 43 38 39 37
jump o¤ 0
1 52 56 56 55 54 46 48 46 43 46
2 51 52 52 53 54 46 48 52 43 44
3 46 54 56 54 52 41 43 46 50 50
4 39 50 52 52 47 44 42 42 43 39
5 33 40 40 45 41 42 43 41 45 37
jump o¤ 1
2 49 52 52 54 57 49 48 52 52 45
3 48 54 56 53 49 43 44 47 42 48
4 45 50 49 49 48 48 45 42 43 43
5 40 42 43 44 41 44 43 39 43 41
jump o¤ 2
3 47 51 51 50 58 45 47 43 50 55
4 51 50 51 50 54 46 48 46 47 50
5 45 43 46 43 41 39 43 41 36 46
jump o¤ 3
4 51 53 53 56 59 50 47 43 47 55
5 48 46 47 44 45 39 44 41 45 48
Notes: Each column, 2–10, reports the share of forecast periods in which the GB forecast
error is larger, ex post, in absolute value than the forecast error from the model labelled at
the top. Entries greater than 50 percent indicate that the model forecast does better than
GB in more than half of forecast periods and are in bold. See also notes to Table 3.
38
Table 5a: RMSPE of de‡ator in‡ation forecasts, 1984-2000
hor. GB RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
jump o¤ -1
0 0.69 1.13 0.96 0.96 0.95 0.93 1.00 1.11 1.80 1.12 0.98
1 0.79 1.31 1.12 1.12 1.08 0.96 1.04 1.35 1.19 1.46 1.09
2 0.81 1.23 1.06 1.05 1.01 0.93 0.96 1.36 1.50 1.55 1.01
3 0.93 1.08 1.15 1.16 1.11 0.96 1.15 1.59 0.97 1.71 1.04
4 0.89 1.16 1.26 1.27 1.21 0.96 1.14 1.66 1.70 1.88 1.07
5 1.14 1.46 1.50 1.54 1.46 1.13 1.38 1.93 1.31 2.80 1.28
jump o¤ 0
1 0.79 0.94 1.00 1.00 0.99 0.96 1.04 1.18 0.99 1.24 0.96
2 0.81 1.06 1.03 1.03 1.00 0.95 0.96 1.30 1.01 1.37 0.97
3 0.93 1.13 1.14 1.13 1.09 1.00 1.12 1.50 0.99 1.55 1.04
4 0.89 0.96 1.16 1.18 1.12 0.94 1.12 1.56 0.88 1.70 0.97
5 1.14 1.21 1.43 1.45 1.39 1.10 1.27 1.83 1.17 2.10 1.19
jump o¤ 1
2 0.81 0.93 0.95 0.95 0.94 0.94 0.94 1.08 0.93 1.16 0.92
3 0.93 1.14 1.11 1.11 1.09 1.00 1.07 1.37 1.12 1.39 1.06
4 0.89 1.13 1.15 1.16 1.12 1.01 1.09 1.47 1.04 1.53 1.03
5 1.14 1.16 1.34 1.35 1.31 1.10 1.28 1.72 1.12 1.82 1.16
jump o¤ 2
3 0.93 0.98 1.06 1.06 1.04 1.02 1.06 1.19 1.01 1.19 1.00
4 0.89 1.06 1.11 1.13 1.10 1.03 1.08 1.33 1.03 1.32 1.20
5 1.14 1.21 1.34 1.35 1.31 1.15 1.29 1.60 1.13 1.58 1.17
jump o¤ 3
4 0.89 0.96 1.05 1.05 1.03 0.99 1.03 1.15 0.97 1.15 0.98
5 1.14 1.22 1.30 1.30 1.28 1.15 1.27 1.48 1.18 1.41 1.19
Notes: Each column, 2–11, reports the root mean square prediction error (RMSPE) for
forecast labelled at top for the current quarter (hor= 0) through 5 quarters into the future.
For forecasts other than Greenbook, the data on which the forecast is based is the real-
time data augmented to bring the data up to the jumping o¤ point using the Greenbook
forecast. Bold indicates that the alternative forecast has smaller RMSPE than Greenbook;
,, indicate that the di¤erence between the RMSPE for the model and Greenbook is
statistically signi…cant at the 1, 5, or 10 percent level, respectively, based on the bootstrap
p-values for the DM test as described in the text.
39
Table 5b: RMSPE of CPI in‡ation forecasts, 1984-2000
hor. GB RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
jump o¤ -1
0 0.49 1.61 1.47 1.47 1.45 1.35 1.46 1.52 1.47 1.45 1.43
1 1.15 1.82 1.48 1.51 1.46 1.35 1.43 1.72 1.57 1.80 1.57
2 1.27 1.70 1.61 1.69 1.65 1.41 1.59 1.65 1.56 1.68 1.57
3 1.56 2.07 1.97 2.08 2.01 1.68 1.92 1.95 1.93 1.91 1.86
4 1.42 2.06 1.85 1.99 1.90 1.48 1.87 2.02 1.86 2.03 1.80
5 1.41 1.93 1.87 1.98 1.91 1.44 1.90 1.98 1.92 2.09 1.76
jump o¤ 0
1 1.15 1.55 1.32 1.32 1.29 1.25 1.26 1.48 1.48 1.47 1.38
2 1.27 1.58 1.49 1.51 1.47 1.31 1.54 1.56 1.44 1.47 1.36
3 1.56 1.61 1.64 1.69 1.65 1.49 1.63 1.68 1.59 1.67 1.52
4 1.42 1.79 1.78 1.85 1.77 1.49 1.69 1.87 1.71 1.85 1.64
5 1.41 1.83 1.79 1.88 1.79 1.41 1.76 1.95 1.73 2.00 1.63
jump o¤ 1
2 1.27 1.38 1.42 1.42 1.41 1.32 1.46 1.41 1.36 1.38 1.28
3 1.56 1.55 1.59 1.60 1.57 1.47 1.59 1.63 1.47 1.65 1.48
4 1.42 1.52 1.61 1.63 1.59 1.42 1.53 1.71 1.51 1.65 1.50
5 1.41 1.60 1.71 1.74 1.68 1.44 1.60 1.81 1.62 1.88 1.56
jump o¤ 2
3 1.56 1.50 1.49 1.49 1.46 1.40 1.43 1.54 1.47 1.53 1.48
4 1.42 1.44 1.56 1.58 1.54 1.40 1.52 1.60 1.42 1.59 1.45
5 1.41 1.44 1.60 1.64 1.59 1.38 1.47 1.66 1.47 1.65 1.47
jump o¤ 3
4 1.42 1.49 1.50 1.50 1.48 1.41 1.47 1.53 1.45 1.49 1.45
5 1.41 1.50 1.55 1.56 1.51 1.37 1.47 1.58 1.45 1.54 1.46
Notes: Each column, 2–11, reports the root mean square prediction error (RMSPE) for
forecast labelled at top for the current quarter (hor= 0) through 5 quarters into the future.
For forecasts other than Greenbook, the data on which the forecast is based is the real-
time data augmented to bring the data up to the jumping o¤ point using the Greenbook
forecast. Bold indicates that the alternative forecast has smaller RMSPE than Greenbook;
,, indicate that the di¤erence between the RMSPE for the model and Greenbook is
statistically signi…cant at the 1, 5, or 10 percent level, respectively, based on the bootstrap
p-values for the DM test as described in the text.
40
Table 5c: RMSPE of output growth forecasts, 1984-2000
hor. GB RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
jump o¤ -1
0 1.75 2.26 1.91 1.91 1.91 1.93 2.27 2.20 2.20 2.30 2.03
1 2.12 2.20 1.85 1.82 1.86 1.97 2.40 2.25 2.29 2.44 1.89
2 2.01 2.49 1.96 1.94 1.90 1.98 2.20 2.06 2.46 2.15 2.20
3 2.15 2.54 2.09 2.04 2.03 2.20 2.35 2.27 2.60 2.18 2.31
4 2.08 2.75 2.05 2.07 2.05 2.18 2.24 2.13 2.56 2.20 2.33
5 2.08 2.97 2.12 2.03 2.01 2.08 2.08 2.08 2.73 2.14 2.53
jump o¤ 0
1 2.12 2.04 1.78 1.78 1.79 1.80 2.04 1.95 2.12 2.11 1.79
2 2.01 2.23 1.88 1.86 1.83 1.80 2.09 1.96 2.20 2.19 2.07
3 2.15 2.36 2.07 2.05 2.04 2.12 2.36 2.23 2.49 2.24 2.19
4 2.08 2.57 2.05 2.02 2.00 2.12 2.18 2.14 2.59 2.23 2.28
5 2.08 2.71 2.12 2.06 2.03 2.17 2.14 2.12 2.59 2.26 2.44
jump o¤ 1
2 2.01 2.31 1.88 1.88 1.83 1.81 1.89 1.85 1.93 1.99 1.91
3 2.15 2.48 2.07 2.06 2.06 2.04 2.29 2.17 2.44 2.25 2.10
4 2.08 2.44 2.05 2.04 2.03 2.06 2.26 2.15 2.40 2.17 2.13
5 2.08 2.54 2.12 2.11 2.09 2.24 2.22 2.17 2.48 2.35 2.25
jump o¤ 2
3 2.15 2.17 2.07 2.07 2.04 1.97 2.07 1.99 2.31 2.09 2.01
4 2.08 2.16 2.03 2.03 2.01 1.96 2.19 2.10 2.32 2.12 2.04
5 2.08 2.21 2.11 2.12 2.11 2.15 2.30 2.18 2.45 2.27 2.12
jump o¤ 3
4 2.08 2.17 2.03 2.03 1.98 1.93 1.98 1.94 2.18 2.02 2.00
5 2.08 2.17 2.08 2.10 2.07 1.99 2.20 2.10 2.31 2.22 2.06
Notes: Each column, 2–11, reports the root mean square prediction error (RMSPE) for
forecast labelled at top for the current quarter (hor= 0) through 5 quarters into the future.
For forecasts other than Greenbook, the data on which the forecast is based is the real-
time data augmented to bring the data up to the jumping o¤ point using the Greenbook
forecast. Bold indicates that the alternative forecast has smaller RMSPE than Greenbook;
,, indicate that the di¤erence between the RMSPE for the model and Greenbook is
statistically signi…cant at the 1, 5, or 10 percent level, respectively, based on the bootstrap
p-values for the DM test as described in the text.
41
Table 6a: Bias and standard deviation of de‡ator in‡ation forecasts
hor. GB RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
bias, 1979–2000
0 0.07 -0.20 0.12 0.12 0.11 -0.02 0.26 0.49 -0.18 0.40 -0.06
1 0.23 -0.06 0.39 0.38 0.37 0.06 0.45 0.86 -0.11 0.65 0.07
2 0.27 0.00 0.45 0.42 0.39 0.07 0.35 1.10 -0.05 0.77 0.13
3 0.35 0.12 0.69 0.71 0.67 0.23 0.53 1.38 0.07 1.07 0.25
4 0.37 0.18 0.84 0.86 0.81 0.31 0.47 1.58 0.16 1.33 0.31
5 0.53 0.33 1.05 1.12 1.04 0.47 0.68 1.86 0.33 1.49 0.46
std. dev., 1979–2000
0 0.81 1.30 1.08 1.08 1.06 1.06 1.16 1.19 1.23 1.24 1.13
1 0.81 1.50 1.20 1.20 1.17 1.17 1.15 1.33 1.42 1.50 1.24
2 0.87 1.27 1.08 1.09 1.06 1.11 1.17 1.23 1.18 1.42 1.18
3 0.97 1.29 1.13 1.15 1.13 1.08 1.39 1.37 1.18 1.42 1.27
4 0.99 1.39 1.18 1.20 1.23 1.16 1.59 1.55 1.42 1.36 1.36
5 1.19 1.68 1.37 1.38 1.40 1.31 1.72 1.57 1.62 1.55 1.60
bias, 1984–2000
0 0.09 -0.15 0.15 0.15 0.13 0.10 0.13 0.38 -0.13 0.58 -0.09
1 0.20 -0.04 0.41 0.41 0.37 0.24 0.31 0.74 -0.02 0.94 0.02
2 0.24 -0.05 0.46 0.44 0.39 0.23 0.23 0.91 0.00 1.06 0.01
3 0.29 -0.02 0.61 0.64 0.58 0.29 0.34 1.12 0.07 1.33 0.04
4 0.32 0.00 0.75 0.77 0.71 0.31 0.36 1.29 0.14 1.61 0.06
5 0.42 0.08 0.92 1.00 0.91 0.37 0.49 1.52 0.28 1.73 0.14
std. dev., 1984–2000
0 0.69 1.12 0.95 0.95 0.94 0.93 1.00 1.05 1.07 0.96 0.98
1 0.76 1.32 1.05 1.05 1.02 0.94 1.00 1.13 1.20 1.13 1.09
2 0.77 1.23 0.96 0.96 0.94 0.91 0.94 1.02 1.05 1.14 1.01
3 0.89 1.08 0.98 0.97 0.95 0.92 1.11 1.14 0.97 1.09 1.04
4 0.84 1.16 1.02 1.02 0.99 0.91 1.09 1.04 1.06 0.97 1.07
5 1.06 1.46 1.19 1.17 1.15 1.07 1.30 1.20 1.28 1.16 1.28
Notes: Each column, 2–11, reports bias or standard deviation of the forecast for the model
labelled at the top. All entries are for jumping o¤ point -1. See also notes to Table 3.
42
Table 6b: Bias and standard deviation of CPI in‡ation forecasts
hor. GB RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
bias, 1979–2000
0 0.13 0.04 0.31 0.31 0.25 -0.10 -0.02 0.18 0.05 0.14 0.18
1 0.27 0.07 0.49 0.48 0.40 -0.16 0.14 0.42 0.14 0.28 0.20
2 0.38 0.23 0.62 0.58 0.51 -0.14 0.20 0.83 0.42 0.62 0.36
3 0.36 0.40 0.97 0.93 0.85 0.06 0.35 1.24 0.71 0.87 0.53
4 0.37 0.39 1.12 1.08 0.98 0.10 0.64 1.42 0.82 0.93 0.52
5 0.54 0.56 1.31 1.31 1.21 0.58 1.02 1.75 1.09 1.17 0.69
std. dev., 1979–2000
0 0.81 2.06 1.73 1.73 1.71 1.62 1.66 1.85 1.94 1.76 1.92
1 1.64 2.21 1.74 1.76 1.74 1.56 1.89 2.12 2.07 1.92 1.91
2 1.75 1.95 1.75 1.87 1.85 1.58 1.95 1.94 1.69 1.84 1.90
3 1.80 2.49 2.14 2.18 2.15 1.87 2.26 2.31 2.15 1.94 2.37
4 1.56 2.55 1.94 1.95 1.96 1.85 2.42 2.38 2.38 2.10 2.31
5 1.55 2.39 1.84 1.78 1.75 2.14 2.21 2.06 2.09 1.85 2.28
bias, 1984–2000
0 0.04 -0.01 0.24 0.24 0.18 -0.07 -0.08 0.06 0.05 0.21 0.04
1 0.10 -0.02 0.45 0.46 0.38 -0.05 0.08 0.23 0.16 0.53 0.03
2 0.19 0.00 0.50 0.54 0.47 -0.07 0.18 0.47 0.34 0.81 0.05
3 0.16 0.03 0.69 0.76 0.66 -0.02 0.28 0.73 0.54 0.96 0.08
4 0.32 0.09 0.93 1.04 0.93 0.14 0.56 0.99 0.75 1.19 0.14
5 0.38 0.14 1.04 1.19 1.07 0.21 0.71 1.21 0.93 1.32 0.19
std. dev., 1984–2000
0 0.49 1.62 1.45 1.45 1.45 1.36 1.47 1.53 1.48 1.44 1.44
1 1.15 1.82 1.42 1.44 1.42 1.35 1.43 1.71 1.57 1.73 1.57
2 1.26 1.71 1.54 1.61 1.59 1.41 1.58 1.58 1.53 1.47 1.57
3 1.56 2.08 1.86 1.95 1.91 1.69 1.90 1.82 1.86 1.65 1.87
4 1.39 2.06 1.61 1.71 1.67 1.48 1.79 1.77 1.71 1.65 1.81
5 1.37 1.93 1.56 1.59 1.59 1.44 1.77 1.57 1.69 1.62 1.75
Notes: Each column, 2–11, reports bias or standard deviation of the forecast for the model
labelled at the top. All entries are for jumping o¤ point -1. See also notes to Table 3.
43
Table 6c: Bias and standard deviation of output growth forecasts
hor. GB RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
bias, 1979–2000
0 -0.41 -0.29 0.13 0.13 0.21 0.02 0.80 0.62 0.85 0.13 -0.10
1 -0.34 -0.04 0.46 0.44 0.55 0.43 1.36 1.05 1.24 0.67 0.16
2 -0.34 -0.23 0.36 0.34 0.43 0.56 1.13 0.89 1.35 0.24 -0.03
3 -0.19 -0.01 0.62 0.48 0.55 0.86 1.20 1.02 1.59 0.33 0.19
4 -0.21 -0.07 0.58 0.56 0.62 0.91 0.94 0.82 1.54 0.33 0.13
5 0.04 0.07 0.73 0.68 0.74 1.02 1.07 0.83 1.62 0.57 0.27
std. dev., 1979–2000
0 2.14 3.30 2.78 2.78 2.71 2.74 2.64 2.67 2.61 2.76 2.75
1 2.74 3.58 2.72 2.73 2.59 2.46 2.46 2.42 2.53 2.59 2.72
2 2.71 3.82 2.81 2.89 2.78 2.81 2.65 2.55 2.65 2.92 2.97
3 2.77 3.41 2.54 2.55 2.49 2.53 2.57 2.56 2.66 2.62 2.84
4 2.57 4.09 2.63 2.66 2.56 2.44 2.39 2.48 2.71 2.66 2.97
5 2.42 3.63 2.57 2.66 2.61 2.68 2.36 2.47 3.25 2.71 2.91
bias, 1984–2000
0 -0.28 0.01 0.13 0.13 0.27 0.29 1.04 0.90 0.89 0.16 -0.13
1 -0.19 0.26 0.38 0.33 0.52 0.74 1.43 1.26 1.23 0.69 0.12
2 -0.14 0.23 0.39 0.32 0.49 0.74 1.37 1.22 1.39 0.29 0.09
3 -0.04 0.47 0.63 0.50 0.65 1.14 1.49 1.35 1.58 0.34 0.32
4 -0.17 0.36 0.53 0.40 0.52 1.04 1.15 1.10 1.42 0.33 0.22
5 -0.09 0.39 0.57 0.39 0.47 0.77 1.00 0.99 1.35 0.35 0.25
std. dev., 1984–2000
0 1.73 2.27 1.91 1.91 1.90 1.91 2.02 2.02 2.03 2.30 2.03
1 2.12 2.19 1.81 1.79 1.79 1.83 1.93 1.87 1.94 2.35 1.89
2 2.01 2.49 1.93 1.92 1.85 1.84 1.73 1.66 2.04 2.14 2.21
3 2.16 2.51 2.00 1.99 1.93 1.88 1.83 1.83 2.08 2.17 2.29
4 2.08 2.74 1.99 2.04 1.99 1.92 1.93 1.83 2.13 2.19 2.33
5 2.09 2.95 2.05 2.00 1.96 1.94 1.84 1.84 2.38 2.12 2.53
Notes: Each column, 2–11, reports bias or standard deviation of the forecast for the model
labelled at the top. All entries are for jumping o¤ point -1. See also notes to Table 3.
44
Table 7: RMSPE of forecasts, most recent vintage
hor. GB RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
de‡ator in‡ation
0 0.66 0.91 0.87 0.87 0.86 0.82 1.00 1.07 0.90 1.05 0.87
1 0.71 1.11 1.06 1.04 1.04 0.92 1.50 1.55 1.09 1.41 1.03
2 0.80 1.11 1.16 1.15 1.15 0.93 1.65 1.81 1.04 1.63 1.12
3 0.89 1.28 1.34 1.35 1.36 1.02 1.80 2.14 1.18 1.80 1.32
4 0.97 1.50 1.52 1.58 1.61 1.09 2.06 2.57 1.39 1.94 1.48
5 1.11 1.57 1.63 1.70 1.72 1.06 2.09 2.86 1.41 1.97 1.57
CPI in‡ation
0 0.77 1.89 1.65 1.65 1.62 1.53 1.68 1.90 1.82 1.62 1.79
1 1.57 2.28 1.96 1.97 1.94 1.72 2.23 2.45 2.06 2.01 2.09
2 1.68 2.06 1.96 2.01 1.99 1.65 2.46 2.58 1.84 2.04 2.00
3 1.78 2.40 2.14 2.20 2.24 1.87 3.41 2.96 2.12 2.26 2.29
4 1.51 2.60 2.19 2.20 2.31 1.99 4.21 3.41 2.36 2.48 2.47
5 1.64 2.58 2.30 2.28 2.36 2.30 4.56 3.58 2.36 2.38 2.52
output growth
0 2.14 3.44 2.99 2.99 2.81 2.74 2.71 2.63 2.90 2.84 3.06
1 3.05 3.86 3.15 3.13 2.95 2.84 3.05 2.95 3.25 3.01 3.26
2 3.01 3.95 2.94 3.00 2.88 2.91 3.05 2.96 3.23 3.16 3.15
3 3.04 4.23 2.89 3.01 2.92 2.98 3.05 2.93 3.48 3.12 3.13
4 2.80 4.46 2.90 3.01 2.90 2.88 3.08 2.88 3.52 2.96 3.14
5 2.63 4.13 2.87 3.04 2.96 2.93 3.12 3.00 3.81 3.03 2.94
Notes: Each column, 2–11, reports the root mean square prediction error (RMSPE) for the
forecast labelled at top for the current quarter (hor= 0) through 5 quarters into the future.
The jumping o¤ point is t1. Other than Greenbook, the forecasts are all based on a single
vintage of data: our most recent vintage. See the notes to Table 3.
45
Table 8: RMSPE of forecasts, Stock-Watson (2005) data
hor. GB RW IAR DAR EWA BMA FAAR FAV IFAV DF UCSV
de‡ator in‡ation
0 0.66 0.91 0.87 0.87 0.85 0.84 0.94 0.94 0.94 1.40 0.87
1 0.71 1.11 1.06 1.04 1.03 1.01 1.16 1.24 1.19 1.50 1.03
2 0.80 1.11 1.16 1.15 1.12 1.05 1.20 1.25 1.17 1.54 1.12
3 0.89 1.28 1.34 1.35 1.28 1.10 1.30 1.37 1.25 1.63 1.32
4 0.97 1.50 1.52 1.58 1.49 1.17 1.47 1.57 1.45 1.61 1.48
5 1.11 1.57 1.63 1.70 1.59 1.18 1.49 1.61 1.49 1.65 1.57
CPI in‡ation
0 0.77 1.89 1.65 1.65 1.57 1.51 1.52 1.81 1.67 1.78 1.79
1 1.57 2.28 1.96 1.97 1.90 1.67 1.81 2.26 1.99 1.96 2.09
2 1.68 2.06 1.96 2.01 1.94 1.56 1.87 1.91 1.89 1.84 2.00
3 1.78 2.40 2.14 2.20 2.12 1.71 2.06 2.09 2.07 2.03 2.29
4 1.51 2.60 2.19 2.20 2.17 1.77 2.22 2.36 2.34 2.08 2.47
5 1.64 2.58 2.30 2.28 2.21 1.84 2.27 2.30 2.36 2.01 2.52
output growth
0 2.14 3.44 2.99 2.99 2.84 2.76 2.59 2.46 2.91 2.73 3.06
1 3.05 3.86 3.15 3.13 2.99 2.90 2.73 2.60 2.84 2.95 3.26
2 3.01 3.95 2.94 3.00 2.97 2.95 3.06 3.00 3.31 3.16 3.15
3 3.04 4.23 2.89 3.01 2.97 2.96 3.07 2.90 3.51 3.11 3.13
4 2.80 4.46 2.90 3.01 2.95 2.84 2.79 2.83 3.54 2.99 3.14
5 2.63 4.13 2.87 3.04 3.02 2.97 2.99 3.07 3.63 2.95 2.94
Notes: Each column, 2–11, reports the root mean square prediction error (RMSPE) for the
forecast labelled at top for the current quarter (hor= 0) through 5 quarters into the future.
The jumping o¤ point is t1. Other than Greenbook, the forecasts are all based on a single
vintage of data: the Stock-Watson (2005) data. See the notes to Table 3.
46
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5
0
0.5
1
1.5
2
2.5
3
Fig. 1: Bootstrap pdf for the ratio of GB to RW RMSPEs in Atkeson−Ohanian sample
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
λ
P
e
r
c
e
n
t
a
g
e

P
o
i
n
t
s
Figure 2: RMSPE of combination forecasts for inflation


h=0
h=1
h=4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
3
λ
P
e
r
c
e
n
t
a
g
e

P
o
i
n
t
s
Figure 3: RMSPE of combination forecasts for growth


h=0
h=1
h=4

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->